ALWAYS_SPLIT seems necessary for ZFP

At least these tests:

```
FAILED tests/ndarray/test_lossy.py::test_lossy[shape1-float64-cparams1-None-False] - RuntimeError: Error while creating the NDArray
FAILED tests/ndarray/test_lossy.py::test_lossy[shape3-float32-cparams3-lossy.b2nd-False] - RuntimeError: Error while creating the NDArray
```

fail if "splitmode" is set to`AUTO_SPLIT`, which is the new default after [this discussion](https://github.com/Blosc/python-blosc2/discussions/337).

Perhaps this is due to the design of the test, but while we can investigate more, they are set explicitly in the test to `ALWAYS_SPLIT`.

Start by running the two failing cases in tests/ndarray/test_lossy.py and inspect how splitmode is configured there. Compare AUTO_SPLIT with the explicit ALWAYS_SPLIT setup, then determine whether the failures come from the test design or the implementation. Done means the cause is established and the affected tests behave correctly with the intended split mode.

Written by the indexing model from the issue text.